Use Seed Evolving with a signed-in browser

To use Seed Evolving with a signed-in browser, run its official API model through a local agent client that can execute tools, then connect that client to Tabbit Browser with Tabbit CLI. Volcengine ModelArk currently names the API model doubao-seed-evolving; the model can propose tool calls, while the local client and CLI do the browser work. The example here copies visible fields from an authorized page into a test form and stops before submission.

The model directory key bytedance-seed/seed-evolving is not the same string as ModelArk's API ID. The official ModelArk model list marks doubao-seed-evolving for function calling and GUI tasks. Availability depends on your region and account activation. No Seed Evolving request was run for this guide, so treat the setup below as a documented workflow to verify in your own client.

Two first-person reports in the Tencent Cloud Developer Community describe client-specific timeouts and manual corrections while building with Seed Evolving. A separate developer account says one feature took three rounds, while another upload issue remained unresolved after a long run. These are coding experiences, not browser tests. They support a modest rule: test your client and keep a person in the loop.

Key takeaways

• Use doubao-seed-evolving with the ModelArk provider. Do not paste a catalog key into the API field without checking it.

• Run the model through a local client that can call both the provider and local tools.

• Use Tabbit CLI to read only the fields needed from the signed-in page.

• Keep a source URL beside each extracted value, then compare the map with the page.

• Save a named draft for review and stop before submitting the form.

Model, client, and browser at a glance

ModelArk documents an OpenAI-compatible endpoint at https://ark.cn-beijing.volces.com/api/v3 and lists clients such as OpenCode, Pi, Codex CLI, and Cline in its tool compatibility guide. That is a supported route to configure; it is not evidence that a specific client configuration or account has been tested. OpenCode lets you define a custom provider and the model's tool capabilities in its V2 provider configuration.

For example, an OpenCode provider can use the official base URL and model ID. Set the key through a local environment variable, and use the current model limits from Ark rather than copying an old catalog value.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "ark/doubao-seed-evolving",
  "providers": {
    "ark": {
      "name": "Volcengine Ark",
      "env": ["ARK_API_KEY"],
      "package": "@opencode/ai/providers/openai-compatible",
      "settings": {"baseURL": "https://ark.cn-beijing.volces.com/api/v3"},
      "models": {
        "doubao-seed-evolving": {
          "name": "Seed Evolving",
          "modelID": "doubao-seed-evolving",
          "capabilities": {"tools": true, "input": ["text", "image"], "output": ["text"]},
          "limit": {"context": 1048576, "output": 262144}
        }
      }
    }
  }
}

Check the exact model and capability fields against the latest Ark model list. OpenCode's custom model entries require you to declare capabilities accurately. The provider documentation also says to keep API secrets out of checked-in configuration.

Ark describes Seed Evolving as a rolling model with a stable API ID. New versions can change behavior without changing the ID. For a repeatable workflow, record the date, API ID, prompt, and tool schema, then rerun a harmless sample after provider updates. See the official Seed Evolving guide.

Connect the local client to the signed-in page

Install Tabbit Browser, sign in to the site yourself, and open the page you are allowed to read. Then install the tabbit skill in the client you will run locally. The Tabbit CLI quick start documents /tabbit for compatible skill commands. Pi uses /skill:tabbit; other clients may expose /skills or a skill picker. Confirm the skill is discoverable, then run tabbit-cli diagnose from the same machine. These are separate checks: finding a skill does not show that the launcher reaches the browser, and a running controller does not show that the target site is signed in.

Use a new named task for this job. Task-owned tabs let you keep ordinary browser tabs available for your own review. The project's local fixture check showed a second task could read the page after a test cookie was set. It displayed Release notes review and GEO-104. That verifies the fixture's Tabbit session flow only; it does not test Seed Evolving, OpenCode, or a third-party account. For setup detail, see the Tabbit CLI guide, existing-browser connection guide, and signed-in session guide.

Copy visible fields into a test form

Ask for a small, read-only extraction first. Give the local client the field names and destination labels. Do not ask it to infer missing values or submit the form.

Use Tabbit CLI in the current Tabbit Browser profile. Read the signed-in source page and map only the visible fields I list to the test form. Keep one row per field with the source label, exact visible value, source page title, and source URL. If a value is absent or ambiguous, leave it blank and add a review flag. Do not click Save, Submit, Send, or any control that changes the record.

Create a browser draft named "Seed Evolving field map — GEO-104 test fixture". Show the field map and wait for my review.

The prompt specifies the action, output, and stopping point. In the suggested test fixture, an expected row could look like this:

This is an illustrative fixture row, not a Seed Evolving model result. For real data, include the exact URL and title the browser shows. Use the first-task checklist to define allowed actions. For choosing a local integration, compare browser MCP with CLI. The Codex setup guide, Tabbit browser overview, and agentic browser explainer cover related setup.

Review the draft before you act

Open the named draft beside the source page. Check every field for exact spelling, dates, units, and status. Reopen each source URL; do not approve a value that the page no longer shows. If a field is unclear, leave it blank and ask a person with access to confirm it.

The practical split is clear: Seed Evolving supplies model output, a local client runs tools, and Tabbit CLI reaches the browser. Start with a harmless page and a test form. If your account cannot select doubao-seed-evolving, do not substitute the directory key; check the current Ark region and account activation first.

Secholar.L

Growth & Content at Tabbit

Secholar.L writes hands-on guides about AI browsers and agent workflows, comparing tools with real research and citing verifiable sources.